Find It or Create It is an AI-driven fashion web platform designed to help users find or create Indian traditional outfits using text or image-based search.
If a user doesn’t find the perfect outfit, they can directly book a designer to stitch a custom design — all within one integrated system.
This project showcases technical and creative expertise across:
This project is part of my portfolio showcase for AI, LLM, and Vision roles.
In India, traditional outfits such as sarees, kurtas, and lehengas vary by region and design.
People often want to find similar outfits from online stores but struggle because image-based searches fail for Indian-specific styles.
To solve this:
A structured dataset was created from scratch to help AI understand and retrieve results efficiently.
Each product record includes:
LEN0001, KUR0002)All data was stored in a single JSON manifest file — acting as the main dataset and connection between AI embeddings and the search interface.
We ensured all data and embeddings remain protected. Secret variables were created in Hugging Face to securely store the dataset and embedding keys, preventing exposure or copying in the public domain.
Each record’s unique ID acts as the link between multiple layers:
This ID-based mapping ensures traceability — for example, if FAISS returns index 158, the corresponding product in JSON can be fetched directly using its mapped ID.
The core AI engine is SigLIP, a Vision Transformer model from Google.
It aligns both images and text in the same vector (embedding) space, enabling cross-modal similarity matching.
Workflow Overview:
.npy files for fast FAISS indexingThis forms a multimodal retrieval system similar to the architecture used in Large Multimodal Models (LMMs) like CLIP and Gemini.
All embeddings are indexed using FAISS (Facebook AI Similarity Search) for high-speed nearest-neighbour lookup.
When user uploads an image:
When user enters a text query:
This structure behaves like Retrieval-Augmented Generation (RAG) — but returns fashion items instead of text outputs.
Here’s how each layer connects (add image of system diagram if needed):
User Input (Text/Image)
↓
Embedding Generation (SigLIP Vision Transformer)
↓
FAISS Similarity Search
↓
Mapping via Unique IDs
↓
Retrieve Data from JSON
↓
Display Results / Book Designer
If a user doesn’t find an ideal match, they can switch to Create It mode for custom stitching.
Gmail confirmation and Telegram alert screenshot
{
"designer_name": "Sofia Anderson",
"designer_id": "DES110",
"customer_name": "Anjali Verma",
"email": "anjali.verma@example.com",
"phone": "+91 9876543xxx",
"location": "Hyderabad, India",
"measurements": {"chest": 34, "waist": 28, "shoulder": 15, "sleeve": 22},
"selected_size": "M",
"search_query": "Indian Traditional Saree with embroidery",
"product_image": "base64_encoded_string"
}
Although SigLIP is a strong Vision Transformer, it’s not trained on Indian traditional fashion datasets.
Hence:
To adapt SigLIP for Indian fashion, we used a hybrid retrieval strategy:
1️⃣ Separate Embedding Pipelines —
We created two FAISS indices:
2️⃣ Localized Matching & Re-Scoring —
After the initial search, we refined the results to make them more accurate and culturally relevant:
Category consistency: Ensures that results match the product type (e.g., a saree search returns only sarees, not kurtas).
Color & keyword matching: Prioritizes results that closely match the color or key terms in the query.
Trusted source preference: Gives higher ranking to products from reliable platforms like Amazon or Flipkart.
This step ensures that the top results are meaningful, accurate, and suitable for Indian traditional fashion.
To improve SEO and result accuracy, we used Ollama (local Mistral LLM) to regenerate product titles into natural, search-optimized queries.
These rewritten queries make it easier to match similar items on Google, Amazon, and Flipkart.
Example:
Input Title: “Women’s Silk Saree with Blouse Piece”
Output Query: “Elegant Silk Saree for Women with Blouse – Traditional Saree”
This step ensures our AI search and online search alignment remain consistent and human-readable.
To make the application publicly accessible and easy to test, the entire system (Flask backend, FAISS index, static UI, and model) was deployed in a Docker container.
Why Docker?
The final Docker image was pushed and deployed on Hugging Face Spaces using custom Docker runtime.
Why Hugging Face Spaces?
| Category | Tools Used | Description |
|---|---|---|
| AI & Embeddings | SigLIP (Vision Transformer) | Image-text alignment for multimodal retrieval |
| Vector Search | FAISS | Fast similarity search across embeddings |
| Backend | Flask | RESTful API and logic layer |
| Data Storage | JSON, NumPy | Lightweight structured mapping |
| Frontend | HTML, CSS, JS | Simple and responsive web interface |
| Optimization | Pillow, Regex | Image resizing & SVG cleanup |
| Deployment | Docker, Hugging Face | End-to-end containerized hosting |
| LLM Integration | Ollama (Mistral) | Query rewriting for better text search |
The current prototype uses around 10,000 image-text pairs — enough for testing, but performance and precision can improve by:
In the future, each product result (Amazon/Flipkart link) can include affiliate tracking.
If a user clicks and purchases, the platform earns a commission-based reward — creating a sustainable revenue model without altering UX.
This project is a portfolio showcase, not a commercial product.
It highlights:
📧 Email: ml.bhanuprakash@gmail.com
🌐 Project Link / Portfolio Showcase: Find It or Create It - AI Fashion Search
🛑 Usage Notice:
This project is for educational and portfolio purposes only.
Reproduction or reuse of the source code without permission is strictly prohibited.
Anyone interested in collaboration or review should contact the author directly.